A Dialogue Analysis Model with Statistical Speech Act
Processing for Dialogue Machine Translation*
Jae-won Lee and Gil Chang
Dept. of Computer Science and CAIR
Korea Advanced Institute of Science and
T e c h n o l o g y , T a e j o n , 305-701, K o r e a
{j wonl ee, gckim}Ocsone, kaist, ac. kr
Abstract
In some cases, to make a proper trans-
lation of an utterance in a dialogue, the
system needs various information about
context. In this paper, we propose a sta-
tistical dialogue analysis model based on
speech acts for Korean-English dialogue
machine translation. T h e model uses syn-
tactic patterns and N-grams reflecting the
hierarchical discourse structures of dia-
logues. The syntactic pattern includes the
syntactic features that are related with the
language dependent expressions of speech
acts. The N-gram of speech acts based
on hierarchical recency approximates the
context. Our experimental results with
trigram showed that the proposed model
achieved 78.59 % accuracy for the top can-
didate and 99.06 % for the top four candi-
dates even though the size of the training
corpus is relatively small. The proposed
model can be integrated with other ap-
proaches for an efficient and robust anal-
ysis of dialogues.
1 Introduction
Recently, special concerns are paid to research on di-
alogue machine translation. Many different aspects
of dialogue, however, make it difficult to translate
spoken language with conventional machine transla-
tion techniques. One of the reasons is that a surface
utterance may represent several ambiguous mean-
ings depending on context. T h a t means such utter-
ance can be translated into many different ways de-
pending on context. Interpreting this kind of utter-
ances often requires the analysis of contexts. There-
fore, the discourse structure of a dialogue plays a
very important role in translating the utterances in
the dialogue. Discourse structures of dialogues are
usually represented as hierarchical structures which
*This research is supported in part by the ministry of
information and communication of Korea.
Kim J u n g y u n Seo
Dept. of Computer Science
Sogang University
Seoul, 121-742, Korea
seo©nlpeng, sogaag, ac. kr
reflect embedding subdialogues (Grosz and Sidner
1986).
Many researchers have studied the way how to
analyze dialogues. One of the representative ap-
proaches is the plan-based method (Litman et al.
1987; Caberry 1989). Considering that our dia-
logue translation system is to be combined with
the speech system to develop an automatic translat-
ing telephone, however, the plan-based approach has
some limitations. In an automatic translating tele-
phone environment, the system must make one cor-
rect translated target sentence for each source sen-
tence and must be able to respond in real time. How-
ever, the plan inference is computationally expensive
and is hard to be scaled up. In order to overcome
such limitations, we have focused on defining mini-
mal approach which uses knowledgebase as small as
possible while it can handle ambiguous utterances.
This paper presents an efficient discourse anal-
ysis model using statistical speech act processing
for Korean-English dialogue machine translation. In
this model, we suggest a probabilistic model which
uses surface syntactic patterns and the N-gram of
speech act reflecting the hierarchical structures of
dialogues to decide the speech act of an input sen-
tence and to maintain a discourse structure. The
proposed model consists of two steps : (1) identify-
ing the syntactic pattern of an utterance (2) calcu-
lating the plausibility for possible speech acts and
discourse relations.
After presenting some motivational examples in
section 2, we discuss the statistical speech act .pro-
cessing model to analyze discourse structure in sec-
tion 3. In section 4, we describe a method to ana-
lyze dialogue structure using the proposed statistical
speech act processing. We discuss experimental re-
sults for the proposed model in section 5. Finally,
we draw some conclusions.
2 Motivation
Translation of dialogues often requires the analysis
of contexts. T h a t is, a surface utterance may be
translated differently depending on context. In this
10
section, we present some motivational examples.
The word 'yey q in Korean has a number of En-
glish expression such as 'yes', 'no', 'O.K:', 'Hello',
'thanks', and so on (Jae-woong Choe 1996). When
the speech act of the utterance 'yey' is response, it
must be translated as 'yes' or 'no'. On the other
hand, when the speech act of the utterance is ac-
cept, it must be translated as 'O.K.'. It is even used
as greeting or opening in Korean. In this case, 'Hello'
is an appropriate expression in English.
The verb 'kulehsupnita' in Korean, also, may
be translated differently depending on context.
Kulehsupnila is used to accept the previous utter-
ance in Korean. In this case, it must be translated
differently depending on context. The following di-
alogue examples show such cases.
Dialogue 1
A : Hankwuk hotelipnikka?
(Is it Hankwuk Hotel?)
B : Yey, kulehsupnita.
(Yes, It is.)
Dialogue 2
A : Kayksil yeyyak hasyesssupnikka?
(Did you reserve a room?)
B : Yey, kulehsupnita.
(Yes, I did.)
To differentiate such cases, a translation system
must analyze the context of a dialogue. Since a dia-
logue has a hierarchical structure than a linear struc-
ture, the discourse structure of a dialogue must be
analyzed to reflect the context in translation. There
are the previous plan-based approaches for analyz-
ing context in dialogues. Since it is very difficult
to have a complete knowledge, it is not easy to find
a correct analysis using such knowledge bases. In
this paper, we propose a statistical dialogue analy-
sis model based on speech acts for dialogue machine
translation. Such model is weaker than the dialogue
analysis model which uses m a n y difference source ,of
knowledge. However, it is more efficient and robust,
and easy to be scaled up. We believe that this kind
of minimal approach is more appropriate for a trans-
lation system.
3 Statistical Speech Act Processing
We construct a statistical dialogue model based on
speech acts as follows.
Let D denote a dialogue which consists of a se-
quence of n utterances, U1, U2 . . . . , Un, and let Si
denote the speech act of Ui. With this notation,
1All notations for Korean follow Yale Romanization
System notation.
11
Sentence Type [ Main-Verb Aux-Verb I Clue-Word I
•
- Assert
YN-Quest t PV
PA
Must
Want t yey
aniyo
WH-Quest FRAG Intent kulemyen
- Imperative LEXEME Possible ...
Serve
Serve_to
May
Intend
Figure 1: A Syntactic Pattern
P(UilUt,U2,...,U i-l) means the probability that
Ui will be uttered given a sequence of utterances
U1,U2,...,Ui-1. As shown in the equation (1),
we can approximate P(UilU1, U2,..., Ui-1) by the
product of the sentential probability P(UilSi) and
the contextual probability P(SilS1,S2,...,Si-1)
(Nagata and Morimoto 1994). In subsequent sec-
tions, we describe the details for each probability.
P(UitU~, U2,..., U,_,) (1)
P(U, IS~)P(S,[S1, $2..... S,_a ).
3.1 Sentential Probability
There is a strong relation between the speaker's
speech act and the surface utterances expressing
that speech act (Allen 1989 ; Andernach 1996). T h a t
is, the speaker utters a sentence which most well
expresses his/her intention (speech act). This sen-
tence allows the hearer to infer what the speaker's
speech act is. However, a sentence can be used as
several speech acts depending on the context of the
sentence.
The sentential probability P(Ui ISi) represents the
relationship between the speech acts and the features
of surface sentences. In this paper, we approximate
utterances with a syntactic pattern, which consists
of the selected syntactic features.
We decided the syntactic pattern which consists
of the fixed number of syntactic features. Sentence
Type, Main- Verb, Aux- Verb, Clue- Word are selected
as the syntactic features since they provide strong
cues to infer speech acts. The features of a syntactic
pattern with possible entries are shown in figure 1.
• Sentence Type represents the m o o d of an ut-
terance. Assert, YN-Quest, WH-Quest, Imper-
ative are possible sentence types.
• Main- Verb is the type of the main verb in the
utterance. PA is used when the main verb rep-
resents a slate and PV for the verbs of type
 Table 1: A part of the syntactic patterns extracted from corpus

Speech Act Sentence Type Main- Verb Aux- Verb Clue Word
Request-Act Imperative PV Request None
Request-Act YN-Quest PV Possible None
Request-Act Assert PV Want None
Ask-Ref WH-Quest PV None None
Ask-Ref YN-Quest PJ None None
Ask-Ref Imperative malhata Request None
Inform Assert PJ None None
Inform Assert PV None None
Request-Conf YN-Quest PJ None None
Request-Conf YN-Quest FRAG None None
Response Assert PJ None yey
Suggest Wh-Quest PV Serve None

event or action. Utterances without verbs be-
long to FRAG (fragment). In the case of per-
formative verbs (ex. promise, request, etc.),
lexical items are used as a Main-Verb because
these are closely tied with specific speech acts.
• Aux-Verb represents the modality such as
Want, Possible, Must, and so on.
• Clue-Word is the special word used in the ut-
terance having particular speech acts, such as
Yes, No, O.K., and so on.
$1, $2, • •., Si-1 were previously uttered. Since pre-
vious speech acts constrain possible speech acts in
the next utterance, contextual information have an
important role in determining the speech act of an
utterance. For example, if an utterance with ask-ref
speech act uttered, then the next speech act would
be one of response, request-conf, and reject. In this
case, response would be the most likely candidate.
The following table shows an example of the speech
act bigrams.

We extracted 167 pairs of speech acts and syntac-

Si--1 Si Ratio
tic patterns from a dialogue corpus automatically ask-ref response 58.46
using a conventional parser. As the result of ap- ask-ref request-confirm 18.46
plying these syntactic patterns to all utterances in
corpus, we found that the average number of speech ask-ref ask-if 7.69
act ambiguity for each utterance is 3.07. Table 1 ask-ref ask-ref 3.08
gives a part of the syntactic patterns extracted from ask-ref suggest 3.08
corpus.
ask-ref inform 1.54
Since a syntactic pattern can be matched with
several speech acts, we use sentential probability,
P(UilSi) using the probabilistic score calculated This table shows that response is the most likely
from the corpus. Equation (2) represents the ap- candidate speech act of the following utterance of
proximated sentential probability. F denotes the the utterances with ask-refspeech act. Also, request-
syntactic pattern and freq denotes the frequency
confirm and ask-if are probable candidates.
count of its argument.
Since it is impossible to consider all preceding ut-
terances $1, $ 2 , . . . , Si-1 as contextual information,
P(U~IS,) ~ P(LIS,) = freq(Fi, Si) (2) we use the n-gram model. However, simply using
fTeq(Si) n utterances linearly adjacent to an utterance as
contextual information has a problem due to sub-
dialogues which frequently occurred in a dialogue.
Let's consider an example dialogue.
3.2 Contextual Probability
The contextual probability P(SilS1, $ 2 , . . . , Si-1) is In dialogue 3, utterances 3-4 are part of an em-
the probability that n utterances with speech act bedded segment. In utterance 3, the speaker asks for
Si is uttered given that utterances with speech act the type of rooms without responding to B's ques-

12
 Dialogue 3

1. A : I would like to reserve a room. request-act

2. B : What kind of room do you want? ask-ref
3. A : What kind of room do you have? ask-ref
4. B : We have single and double rooms. response
5. A : A single room, please. response

tion (utterance 2). This subdialogue continues up

to the utterance 4. As shown in the above exam-
ple, dialogues cannot be viewed as a linear sequence
of utterances. Rather, dialogues have a hierarchical request- ~acce
structure. Therefore, if we use n utterances linearly
adjacent to an utterance as a context, we cannot re-
RA = © action pt/e

fleet the hierarchical structure of a dialogue in the

model. JRC/
Therefore, we approximate the context for an ut-
terance as speech acts of n utterances which is hi-
erarchically recent to the utterance. An utterance RI/
A is hierarchically recent to an utterance B if A is ask-if/ Q~respo
adjacent to B in the tree structure of the discourse RI =
O ask-ref _ nse _ O
(Walker 1996). Equation (3) represents the approxi-
mated contextual probability in terms of hierarchical
recency in the case of using trigram. In this equa-
tion, Ui is adjacent to Uj and Uj is adjacent to Uk
in the discourse structure, where 1 _< j < k _< i - 1 . request- R ~ c o n fi
RC :
O confirm ~
rm/e O
P(SilS1,S2 .... , & - l ) '~' P(SilSj,S~). (3)

4 Discourse Structure Analysis inform/

response
Now we can define a discourse structure analysis
model with the statistical speech act processing.
GI :
(3
Formally, choose Si which maximizes the following RI/ -~"-k...J
probability
Figure 2: A part of the dialogue transition network
max P( F~IS~)P( S~ISj, Sk ). (4)
S,

where Si is a possible speech act for the utterance

Ui. Uj and Uk are the utterances which Uj is hi- A:request
erarchically adjacent to Ui, and Uk to Uj, where 1 acti°n~t'--)
<_j<k<_i-1. RA-I : v ;i
In equation (4), one problem is to search all pos-
sible Uj that Ui can be connected to. We use the
dialogue transition networks (DTN) and a stack for
maintaining the dialogue state efficiently. The di-
alogue transition networks describe possible flow of
speech acts in dialogues as shown in figure 2 (Seo
et al. 1994, Jin Ah Kim et al. 1995). Since DTN k-ref ~ O B:response~
is defined using recursive transition networks, it can RI-II u3 u4

handle recursively embedded subdialogues. It works

just like the RTN parser (Woods 1970). If a subdi-
alogue is initiated, a dialogue transition network is Figure 3: The transitions of dialogue 3
initiated and a current state is pushed on the stack.
On the other hand, if a subdialogue is ended, then a

13
dialogue transition network is ended and a current
state is popped from the stack. This process contin- Table 2: The distribution of speech acts in corpus
ues until a dialogue is finished.
With DTN and the stack, the system makes ex-
pectations for all possible speech acts of the next Speech Act Type Ratw Speech Act Type Ratio
utterance. For example, let us consider dialogue 3. ask-ref ask-if
Figure 3 shows the transitions with the dialogue 3. inform response
In utterance 2, according to the RA diagram
request-confirm request-action
in figure 2, B may request-confirm or request-
il~formation. Since B asks for the type of rooms, suggest confirm
push operation occurs and a RI diagram is initi- accept reject
ated. In utterance 3, A doesn't know the possible correct promise
room sizes, hence asks B to provide such informa- expressive greeting
tion. Therefore, push operation occurs again and a
good-bye Total
new RI diagram is initiated. This diagram is con-
tinued by response in utterance 4. In utterance 5,
this diagram is popped from the stack by response
for ask-refin utterance 2.
Table 3: Experimental results
In this state, some cases can be expected for the
next utterance. The first case is to clarify the ut- 1 2 3 4
terance 5. The second case is to return to the ut-
Model I 68.48 % 74.57 % 76.09 % 76.30 %
terance 1. The last case is to introduce a new sub-
dialogue. Therefore, if we assume that ask-if and Model II 78.59 % 92.82 % 97.88 % 99.06 %
request-confirm are possible from the syntactic pat-
tern of the next utterance, then the following table
can be expected for the next utterance from the di-
alogue transition networks. /SP/hotel
IKSlEtten pangul wenhasipnikka?
/ES/What k i n d o f room do you want?
Uk Uj U~ /SA/ask-ref
/DS/[1]
(0:-:init) (0:-:init) (6:B:ask-if)
(2:B:ask-ref) (5:A:response) (6:B:ask-if)
/SP/customer
(2:B:ask-ref) (5:A:response) (6:B:request-conf) IKSIEtten pangiisssupnikka?
(0:-:init) (l:A:request-act) (6:B:ask-if) /ES/What k i n d of room do you have?
/SA/ask-ref
/DS/[1,1]
Since DTN has the same expressive power as
ATN(Augmented Transition Network) grammar, we
believe that it is not enough to cover the whole phe- We test two models in order to verify the efficiency
nomenon of dialogues. However, considering the fact of the proposed model. Model-I is the proposed
that the utterances requiring context for translation model based on linear recency, where an utterance U/
is relatively small, it is practically acceptable for di- is always connected to the previous utterance Ui-1.
alogue machine translation. Model-II is the model based on hierarchical recency.
Table 3 shows the average accuracy of two models.
5 Experiments and Results Accuracy figures shown in table 3 are computed
by counting utterances that have a correct speech
In order to experiment the proposed model, we used act and a correct discourse relation. In the closed
70 dialogues recorded in real fields such as hotel experiments, Modelq achieved 68.48 % accuracy for
reservation and airline reservation. These 70 dia- the top candidate and 76.30 % for the top four can-
logues consist of about 1,700 utterances, 8,319 words didates. In contrast, the proposed model, Model-
total. Each utterance in dialogues was annotated II, achieved 78.59 % accuracy for the top candidate
with speech acts (SA) and with discourse structure and 99.06 % for the top four candidates. Errors in
information (DS). DS is an index that represents the Model-I occurred, because the hierarchical structure
hierarchical structure of discourse. Table 2 shows of dialogues was not considered. Although dialogue
the distribution of speech acts in this dialogue cor- corpus are relatively small, the experimental results
pus. The following shows a part of an annotated showed that the proposed model is efficient for ana-
dialogue corpus. lyzing dialogues.

14
6 Conclusions
In this paper, we described an efficient dialogue anal-
ysis model with statistical speech act processing. We
proposed a statistical method to decide a speech
act of a sentence and to maintain a discourse struc-
ture. This model uses the surface syntactic patterns
of the sentence and N-gram of speech acts of the
sentences which are discourse structurally recent to
tile sentence. Our experimental results with trigram
showed that the proposed model achieved 78.59 %
accuracy for the top candidate and 99.06 % for the
top four candidates although the size of the train-
ing corpus is relatively small. This model is weaker
than the dialogue analysis model which uses many
difference source of knowledge. However, it is more
efficient and robust, and easy to be scaled up. We
believe that this kind of statistical approach can be
integrated with other approaches for an efficient and
robust analysis of dialogues.
References
Hwan Jin Choi, Young Hwan Oh, 1996, "Analysis of
Intention in Spoken Dialogue Based on Learning
of Intention Dependent Sentence Patterns", Jour-
nal of Korea Science Information Society, Vol.23,
No.8, pp.862-870, In Korea.
Jae-woong Choe, 1996, "Some Issues in Conversa-
tional Analysis : Telephone Conversations for Ho-
tel Reservation," In Proc. of Hangul and Korean
Language Information Processing, pp.7-16, In Ko-
rea.
James F. Allen, C. Raymond Perrault, 1980, "Ana-
lyzing Intention in Utterances", Artificial Intelli-
gence, Vol.15, pp.143-178
Elizabeth A. Hinkelman, James F. Allen, 1989,
"Two Constraints on Speech Act Ambiguity," In
Proc. of th 27th Annual Meeting of the ACL, As-
sociation of Computational Linguistics, pp.212-
219.
Barbara J. Grosz, Candace L. Sidner, 1986, "Atten-
tion, Intentions, and the Structure of Discourse",
Computational Linguistics, Vol.12, No.3, pp.175-
204.
Philip R. Cohen, C. Raymond Perrault, 1979, "El-
ements of a Plan-Based Theory of Speech Acts",
Cognitive Science, Vol.3, pp.177-212.
Diane J. Litman, James F. Allen, 1987, "A Plan
Recognition Model for Subdialogues in Conversa-
tions", Cognitive Science, Vol.11, pp.163-200.
Hiroaki Kitano, 1994, "Speech-to-Speech Transla-
tion : A Massively Parallel Memory- Based Ap-
proach",Kluwer Academic Publishers.
Jan Alexandersson, Elisabeth Maier, Nobert Rei-
thinger, 1994, "A Robust and Efficient Three-
15
Layered Dialogue Component for a Speech-to-
Speech Translation System", Proc. of the 7th Eu-
ropean Association for Computational Linguis-
tics, pp. 188-193.
Jin Ah Kim, Young Hwan Cho, Jae-won Lee, Gil
Chang Kim, 1995, "A Response Generation in
Dialogue System based on Dialogue Flow Dia-
grams," Natural Language Processing Pacific Rim
Symposium, pp.634-639.
Jungyun Seo, Jae-won Lee, Jae-Hoon Kim, Jeong-
Mi Cho, Chang-Hyun Kim, and Gil Chang Kim,
1994, "Dialogue Machine Translation Using a Di-
alogue Model", Proc. of China-Korea Joint Sym-
posium on Machine Translation, pp.55-63.
Masaaki Nagata and Tsuyoshi Morimoto, 1994,
"First Steps towards Statistical Modeling of Di-
alogue to Predict the Speech Act Type of the
Next Utterance", Speech Communication, Vol.15,
pp.193-203.
Massko Kume, Gayle K. Sato, Kei Yoshimoto,
1990, "A Descriptive Framework for Translat-
ing Speaker's Meaning", Proc. of the 4th Euro-
pean Association for Computational Linguistics,
pp.264-271.
Marilyn Walker and Steve Whittaker, 1990, "Mixed
initiative in Dialogue : An Investigation into Dis-
course Segmentation", In Proc. of the 28th An-
nual Meeting of the ACL, Association of Compu-
tational Linguistics, pp.70-78.
Sandra Caberry, 1989, "A Pragmatics-Based Ap-
proach to Ellipsis Resolution", Computational
Linguistics, Vol.15, No.2, pp.75-96.
Toine Andernach, 1996, "A Machine Learning Ap-
proach to the Classification of Dialogue Utter-
ances", Proceedings of NeMLaP-2, Bilkent Uni-
versity, Turkey.
Marilyn A. Walker, 1996, "Limited Attention and
Discourse Structure,", Computational Linguistics,
Vol.22, No.2, pp.255-264.
Woods, W. A., 1970, "Transition Network Gram-
mars for Natural Language Analysis," Commun.
of the ACM, Vol.13, pp.591-606.